{
  "id": 16002433,
  "title": "NVIDIA Cosmos Opens New Tools to Train, Evaluate and Deploy World Models",
  "url": "https://urgent.news/2026/10/22/nvidia-cosmos-opens-new-tools-to-train-evaluate-and-deploy-world",
  "topic": "ai",
  "section": "AI",
  "published": "2026-10-22T05:00:27.000Z",
  "source": {
    "name": "NVIDIA Blog",
    "slug": "nvidia-blog",
    "url": "https://blogs.nvidia.com/blog/cosmos-open-train-evaluate-deploy-world-models/"
  },
  "original_language": "en",
  "account": "NVIDIA has rolled out additional tools within the Cosmos platform to streamline the creation, assessment, and deployment of world models, which enable physical AI systems to comprehend and simulate the real world. The company's platform leads the pack on six different open models leaderboards, providing world model developers with access to frontier open models, data, training and optimization tools, recipes, and blueprints.\n\nAccelerating the process of creating training data, the NVIDIA Cosmos Data Factory provides a suite of cloud-native blueprints. These include Autonomous Vehicle Dataset Generation, which assembles and labels driving data while incorporating synthetic examples of difficult-to-capture scenarios; Image Attribute Augmentation, which varies object appearance to enhance image search results; Event Video Generation, converting static images into synthetic videos of rare or hazardous events; Defect Image Generation, producing printed circuit board, metal, and glass defect images for inspection models; and Video Data Augmentation, altering weather and lighting in existing footage while maintaining scene structure.\n\nCosmos Data Factory's blueprints are available as GitHub projects on Microsoft Azure, Nebius, and CoreWeave. Porsche is leveraging Cosmos Data Factory through Voxel51 on Microsoft Azure to construct a shared autonomous vehicle data flywheel spanning Volkswagen Group brands, aiming to cut repeated physical test drives by up to 90% and reduce development cycle times by up to 85%. Additionally, ANYbotics used Cosmos Data Factory blueprints to generate over 1,500 synthetic video clips, validating outputs with Encord's platform for training its industrial-inspection robots to spot liquid spills and other changes between facility visits. Ultralytics improved defect localization and recall by retraining its YOLO model with real images and targeted synthetic defects generated using Cosmos Data Factory.\n\nDevelopers can train their chosen architecture or start with a pretrained model in Cosmos Framework. Their options include fine-tuning and reinforcement learning on open models like GR00T, Alpamayo, QWEN, Pi0.5, and other Hugging Face models compatible with Cosmos Framework. They can also optimize models with distillation and quantization and deploy with top-tier inference performance. MicroAGI, a robotics research lab, fine-tuned Cosmos 3 on videos of humans manipulating objects and distilled its 32-billion-parameter reasoner to an 8-billion-parameter Nano model without a loss in accuracy. Developers can label data, post-train Cosmos 3 on Roboflow for their specific environments and tasks, and then deploy these models to RoboFlow hosted or on-premises NVIDIA GPUs to bring video reasoning to factories and warehouses.",
  "summary": "To help developers rapidly build world models, which help physical AI systems understand and simulate the real world, NVIDIA has expanded the Cosmos platform with new training tools, simulation checkpoints and Cosmos Data Factory blueprints. Ranking top five on six different open models leaderboards, the NVIDIA Cosmos platform gives world model builders the frontier open […]",
  "key_points": [],
  "editors_take": null,
  "illustration": null,
  "coverage": {
    "outlets": 1,
    "also_reported_by": []
  },
  "ai_generated": true,
  "disclaimer": "Summaries, key points and the editor’s take are written by software from other outlets’ reporting and may contain errors — always check the linked original."
}